search

TheSouthFrog / stylealign

[ICCV 2019]Aggregation via Separation: Boosting Facial Landmark Detector with Semi-Supervised Style Transition
183 stars 28 forks source link

Question about the performance #19

Open flyingmrwang opened 4 years ago

flyingmrwang commented 4 years ago

Hi, this is a great work. After study, I have several questions:

  1. If I wanna improve my model with your augmentation, can I just download your augmented dataset and do the training? Furthermore, is it ok to combine origin images with stylized ones?
  2. The images are heavily stylized, which has a gap between images in real life. Will this perform better than just train with origin images?
  3. The augmented images has unclear contour, will this cause inaccurate contour points prediction?